In this paper we present a method based on SVMs by regularized risk minimization for the facial feature selection aiming at improving performance of the classifier by (1) using WT + KPCA as filter approach to choose a set of more meaningful representatives to replace the original data for feature selection; (2) using SVM RFE iterative procedure as wrapper approach to obtain the optimum feature subset; (3) using regularized risk minimization as feature selection ranking criterion. Experimental results on FERET face database subsets indicate that the proposed method has a significant improvement in the classification accuracy and speed.